Europe’s largest cloud provider no longer wants to only rent out the computers that train AI. It wants to build the AI itself. OVHcloud frontier AI models are the French company’s plan to train its own large language models from scratch and become a home-grown challenger to Mistral AI, Europe’s best-known model maker.
CEO Octave Klaba set out the plan at the VivaTech conference in Paris on 17 June 2026. He said a project that once cost around €1 billion can now be done for €150-200 million, Reuters reported.
Key takeaways
- OVHcloud plans to train frontier AI models (the most advanced kind, built from scratch rather than adapted from someone else’s).
- It has already finished pre-training one model on JUPITER, Europe’s fastest supercomputer.
- Klaba puts the cost at €150-200 million, down from about €1 billion a few years ago.
- The company says it will not train on customer data and wants to open-source the models once they are good enough.
- No benchmarks or launch date had been published at the time of the announcement.
What exactly did OVHcloud announce?
The short version: a family of models, not a single chatbot. Klaba argued that rivals now ship several models, each tuned for a job, so OVHcloud will do the same. Think of a toolbox rather than one Swiss Army knife.
The first model has finished pre-training. That is the expensive first stage, where a model reads huge amounts of text and learns the patterns of language. Fine-tuning and safety work come after. The original Reuters headline described OVHcloud as aiming to become Europe’s second LLM player, a label that was later softened. Either way, the target is clear: sit next to Mistral as a European option to American and Chinese AI.
Why can a cloud company afford this now?
Klaba gave three reasons for the cost drop: cheaper and better chips, smarter training methods, and synthetic data, meaning text generated by existing AI to help train new AI. Together, he said, they shrink a €1 billion bill to as little as €150 million.
He called it a “second wave” of AI, where newcomers build on the path OpenAI, Anthropic and Mistral already cut. Those are his estimates, not audited numbers. Still, the logic holds: the second climber up a mountain spends less because the route is already mapped.
The compute also came from outside. JUPITER, run by Forschungszentrum Jülich in Germany under the EuroHPC programme, opened in June 2025 as Europe’s first exascale system, able to do a billion billion calculations per second. It ranked fourth in the world on the TOP500 list when it opened.
Where did OVHcloud get the AI know-how?
It bought it. On 25 March 2026, OVHcloud signed a binding agreement to acquire Dragon LLM, a Paris startup founded as Lingua Custodia to serve the finance sector. Dragon LLM builds specialised models for regulated industries that can run on a customer’s own servers, and it won the European Commission’s Large AI Grand Challenge.
The deal, whose price was not disclosed, launched OVHcloud’s AI Lab. Its goal is to train and fine-tune “sovereign” LLMs, meaning models built, hosted and governed in Europe under European law.
OVHcloud was not starting from zero on the customer side either. In April 2025 it launched AI Endpoints, a pay-as-you-go service offering API access to more than 40 open models, including Llama and Mistral variants. Its own models would slot into that shop window.
OVHcloud vs Mistral: how do they compare?
The two are partners as well as rivals: Mistral models already run on OVHcloud’s platform. Here is how their positions differed as of June 2026.
| Factor | OVHcloud | Mistral AI |
|---|---|---|
| Core business | Cloud hosting and data centres | Building AI models |
| Model status | First model pre-trained, no public release | Several models already released |
| Strategy | Family of specialised models | Mix of open and commercial models |
| Training data promise | No customer data used | Not part of this announcement |
| Open source | Planned once performance is good enough | Some models released with open weights |
OVHcloud’s real edge is distribution. It already owns the servers and the customer contracts, so it does not have to rent compute from an American hyperscaler to serve its models.
Why does Europe want its own AI models?
Governments and companies are looking for options that do not depend on US or Chinese suppliers. If your hospital, bank or ministry runs on an AI model, you want to know who controls it, where the data sits, and whether it can be switched off by a foreign decision. The same instinct sits behind moves such as the US push to pull DeepSeek from federal agencies, just pointed in the other direction.
Dependence on one provider is also a reliability risk. The Revolut outage that took down 31 services showed how a single cloud link can ripple out to millions of users.
My take, clearly as opinion: a cloud company that builds models makes more sense than a model lab that has to buy cloud. The open question is whether OVHcloud can hire and keep the research talent that Mistral, Google and OpenAI fight over.
What happens next
- Benchmarks. OVHcloud had not published any performance figures by the time of the announcement. Independent test results will show whether “frontier” is earned or aspirational.
- Open-source timing. Klaba said the company would decide when the models are good enough to release openly. Watch for the licence terms.
- Product integration. Expect the models to appear first inside AI Endpoints and Dragon LLM’s on-premises tools for regulated industries.
- Funding. €150-200 million is a large sum for a cloud company. How it is financed will matter to OVHcloud’s investors.
FAQ
Has OVHcloud released its frontier AI model yet?
Not as of the June 2026 announcement. One model had finished pre-training on the JUPITER supercomputer, but no release date or benchmarks were given.
Will OVHcloud use customer data to train its models?
No. CEO Octave Klaba said customer data will not be used to train the models.
How much will OVHcloud spend on its AI models?
Klaba estimated €150-200 million, compared with about €1 billion for a similar project a few years earlier. These are company estimates, not audited figures.
If your team handles sensitive data in Europe, add OVHcloud to your AI shortlist now and check its benchmarks as soon as they are published.
Leave a comment